An Empirical Study on the Developers Perception of Software Coupling

Transcription

1 An Empirical Study on the Developers Perception of Software Coupling Gabriele Bavota Bogdan Dit Rocco Oliveto Massimiliano Denys Di Penta Poshyvanyk Andrea De Lucia

2 Coupling Module 1 Module 2

3 Coupling Module 1 Module 2

4 Coupling Measure coupling Module 1 Module 2

5 Coupling Measure coupling Use coupling - predicting fault proneness - change impact analysis - software remodularization - software reuse - change propagation -etc. Module 1 Module 2

6 What are the Typical Types of Information used for Coupling?

7 What are the Typical Types of Information used for Coupling? [Briand et al., TSE 99]: overview of structural metrics

8 What are the Typical Types of Information used for Coupling? [Briand et al., TSE 99]: overview of structural metrics Typical types of information: Structural Dynamic Semantic Logical

9 Structural

10 A attribute_a getattributea() setattributea( ) methoda1() methoda2()

11 A attribute_a getattributea() setattributea( ) methoda1() methoda2() Methods sharing attributes are more likely to implement similar responsibility

12 A attribute_a getattributea() setattributea( ) methoda1() methoda2() Methods grouped together by developers are more likely to implement similar responsibility

13 B A attribute_a getattributea() setattributea( ) methoda1() methoda2() Classes having inheritance relations are more likely to implement similar responsibility

14 B A attribute_a getattributea() setattributea( ) methoda1() methoda2() Calls between methods (Classes) C methodc1() methodc2()

15 D attribute_d methodd1() methodd2() methodd3() B A attribute_a getattributea() setattributea( ) methoda1() methoda2() E methode1() Calls between methods (Classes) C methodc1() methodc2()

16 Structural Dynamic

17 B D attribute_d methodd1() methodd2() methodd3() A attribute_a getattributea() setattributea( ) methoda1() methoda2() E methode1() C methodc1() methodc2()

18 B D attribute_d methodd1() methodd2() methodd3() A attribute_a getattributea() setattributea( ) methoda1() methoda2() E methode1() C methodc1() methodc2() Execution Trace A.methodA2() C.methodC2() C.methodC2() E.methodE1()

19 B D attribute_d methodd1() methodd2() methodd3() A attribute_a getattributea() setattributea( ) methoda1() methoda2() E methode1() C methodc1() methodc2() Execution Trace A.methodA2() C.methodC2() C.methodC2() E.methodE1()

20 The calls that are not in the trace will not be used for computing dynamic coupling D attribute_d methodd1() methodd2() methodd3() B A attribute_a getattributea() setattributea( ) methoda1() methoda2() E methode1() C methodc1() methodc2() Execution Trace A.methodA2() C.methodC2() C.methodC2() E.methodE1()

21 The calls that are not in the trace will not be used for computing dynamic coupling D attribute_d methodd1() methodd2() methodd3() B A attribute_a getattributea() setattributea( ) methoda1() methoda2() E methode1() Dynamic coupling is more precise than structural coupling C methodc1() methodc2() Execution Trace A.methodA2() C.methodC2() C.methodC2() E.methodE1()

22 The calls that are not in the trace will not be used for computing dynamic coupling D attribute_d methodd1() methodd2() methodd3() B A attribute_a Dynamic coupling getattributea() C is harder to generate than setattributea( ) structural coupling methoda1() methodc1() methoda2() methodc2() E methode1() Dynamic coupling is more precise than structural coupling Execution Trace A.methodA2() C.methodC2() C.methodC2() E.methodE1()

23 Structural Dynamic Semantic

24

25 Remove special characters

26 Remove special characters Remove common words

27 Remove special characters Remove common words Split identifiers

28 Compute textual similarity between Textual documents coupling

29 Structural Dynamic Semantic Logical (Evolutionary)

30 Source File A B C D E Time

31 Source File Changes to file A A B C D E Time

32 Source File A B C D E Time

33 Source File A, D & E changed together often A B C D E Time

34 Source File B & C changed together often A B C D E Time

35 Source File B & C changed together often A B C D Classes E that often change together are likely to implement similar responsibility Time

36 Which type of coupling aligns better with developers perception?

37 Which type of coupling aligns better with developers perception? Empirical study with 76 developers

38 Choosing Software Systems

39

40 Size 200 KLOC 29 KLOC 28 KLOC # of classes 1, Vocabulary 7,961 2,834 2,418

41 Size 200 KLOC 29 KLOC 28 KLOC # of classes 1, Vocabulary 7,961 2,834 2,418

42 Size 200 KLOC 29 KLOC 28 KLOC # of classes 1, Vocabulary 7,961 2,834 2,418 Historical 8 years 5 years 12 years

43 Size 200 KLOC 29 KLOC 28 KLOC # of classes 1, Vocabulary 7,961 2,834 2,418 Historical 8 years 5 years 12 years Execution Traces Coverage

44 No test cases available ArgoUML

45 No test cases available ArgoUML Trace 1

46 No test cases available ArgoUML Trace 1 Trace 2

47 No test cases available Trace i ArgoUML Trace 1 Trace 2

48 Size 200 KLOC 29 KLOC 28 KLOC # of classes 1, Vocabulary 7,961 2,834 2,418 Historical 8 years 5 years 12 years Execution Traces Coverage 66% 67% 86%

49 Choosing Software Systems Choosing Participants

50 Original Developers

51 Original 6 Developers 3 3

52 Original 6 Developers 3 3 External Developers 64 64

53 Original 6 Developers 3 3 External Developers with industrial experience 7 faculty 16 PhD students 33 MS students 3 undergraduates

54 Original 6 Developers 3 3 External Original developers evaluated Developers their own system

55 Developers External Developers External developers evaluated all systems

56 Choosing Software Systems Choosing Participants Coupling Metrics

57

58 Structural Coupling: ICP: Information flow-based coupling [Lee et al. 95] Dynamic Coupling: IC_CD: Import Coupling Class Dynamic Message [Arisholm et al. 04] Semantic Coupling: CCBC: Conceptual Coupling Between Classes [Poshyvanyk et al. 08] Logical Coupling: Association rule-based Change Coupling [Ying et al. 04] [Zimmerman et al. 04]

59 Choosing Software Systems Choosing Participants Coupling Metrics Data Collection

60 Structural

61 Structural

62 Structural High Structural Coupling: - Top 2 pairs

63 High Structural Coupling: - Top 2 pairs Structural Low Structural Coupling: - Bottom 2 pairs

64

65 No coupling Highly coupled Likert Scale

66 Structural Dynamic Semantic Logical

67 Structural Dynamic 16 pairs of classes / system Semantic Logical

68 Choosing Software Systems Choosing Participants Coupling Metrics Data Collection Results

69 Likert scale for coupling perceived by developers

70 Types of high coupling

71 Types of high coupling If developers agree with high coupling, results should be here

72 If developers agree with low coupling, results should be here Types of low coupling

73 External Dev. Original Dev.

74 Statistical analysis External Dev. Original Dev.

75

76 64 External Dev. 6 Original Dev.

77 64 External Dev. 6 Original Dev. Classes with high semantic measure are perceived as strong coupled

78 Semantic outperforms others [p-value<0.05] 64 External Dev. 6 Original Dev.

79 64 External Dev. 6 Original Dev. Structural and dynamic align well with developers perception

80 Logical is outperformed by others 64 External Dev. 6 Original Dev.

81 Qualitative example 64 External Dev. 6 Original Dev.

82 ActionVisibilityPrivate ActionVisibilityProtected Original developer: these classes have a very high coupling even if there is no cooperation between them

83 Semantic measure is the only one capturing coupling between these classes Original developer: these classes have a very high coupling even if there is no cooperation between them

84 64 External Dev. 6 Original Dev. Classes with low semantic measure are perceived as loose coupled

85 Semantic outperforms others [p-value<0.05] 64 External Dev. 6 Original Dev.

86 Wider distribution of values why? 64 External Dev. 6 Original Dev.

87 AboutBox FigCompartmentBox Classes with low dynamic coupling, but ranked as strong coupled.

88 AboutBox FigCompartmentBox Original developer: Classes with low dynamic coupling, but ranked as strong coupled. even if indirectly, the two classes exhibit some form of coupling since they implement similar jobs and changes to the GUI are likely to involve both classes

89

90 64 External Dev. 3 Original Dev.

91 64 External Dev. 3 Original Dev. Classes with high semantic measure are perceived as strong coupled

92 Semantic outperforms others [p-value<0.05] 64 External Dev. 3 Original Dev.

93 Structural is most unsuited to capture coupling 64 External Dev. 3 Original Dev.

94 64 External Dev. 3 Original Dev. Dynamic and semantic are the only ones that properly capture low coupling

95 Structural produces strange results Why? 64 External Dev. 3 Original Dev.

96 ZoomTool ZoomUpdateStrategy Classes with low coupling => perceived as coupled

97 ZoomTool ZoomUpdateStrategy Classes with low coupling => perceived as coupled Original developer: these classes are peer features participating in the same context

98 ZoomTool ZoomUpdateStrategy Classes with low coupling => perceived as coupled Original developer: these classes are peer features participating in the same context Semantic measure captures this coupling (CCBC=0.52)

99 Semantic measure is not a silver bullet AWTCursor Locator Not coupled, but semantic measure is high (CCBC=0.41)

100

101 64 External Dev. 3 Original Dev.

102 Same trends as other systems 64 External Dev. 3 Original Dev.

103 How can these types of coupling affect the remodularization of a software system?

104 Original software structure Package 1 Package 2 A C B F D G E Package n H I

105 Original software structure Package 1 Package 2 A C B F D G E Package n H I A B D I C G E D Structural dependencies

106 Original software structure Package 1 Package 2 A C B F D G E Package n H I A B D I C G E D Structural dependencies Bunch [Mitchell and Mancoridis 06]

107 Original software structure Package 1 Package 2 A C B F D G E Package n H I A C B D I G E D Package i A D B H Package j F I C E Structural dependencies Bunch [Mitchell and Mancoridis 06]

108 Original software structure Package 1 Package 2 A C B F D G E Package n H I A C B D I G E D Package i A D B H Package j F I C E Structural dependencies Bunch [Mitchell and Mancoridis 06]

109 Original software structure Package 1 Package 2 A C B F D G E Package n H I A C MoJoFM [Wen and Tzerpos 04] B D I G E D Package i A D B H Package j F I C E Structural dependencies Bunch [Mitchell and Mancoridis 06]

110 Original software structure Package 1 Package 2 A C B F D G E Package n H I How many move and A C join operations are required B Gto transform the produced D I E modularization to the original one? D Structural dependencies Package i A D MoJoFM [Wen and Tzerpos 04] B H Package j Bunch [Mitchell and Mancoridis 06] F I C E

111 Median MoJoFM for 30 runs of Bunch (to account for randomness)

112 Median MoJoFM for 30 runs of Bunch (to account for randomness) Structural and semantic coupling produce better remodularity

113 Median MoJoFM for 30 runs of Bunch (to account for randomness) Maybe not enough data?

114 Threats to Validity Metrics frequently used in the literature Performance of dynamic and logical metrics may have been impacted by insufficient data Pairs of classes may not be representative Remodularization task: use of current modularization of the system as oracle Choice of coupling measures

115 Conclusions Empirical study to assess the developer perception on software coupling 76 developers Investigate which type of information is more suitable for remodularization We need more studies to generalize the results

116 Future Work Consider systems with more dynamic and logical information Use different coupling measures Consider combinations of coupling metrics

117 Thank you! Questions?

118 References B. S. Mitchell and S. Mancoridis, On the automatic modularization of software systems using the bunch tool, IEEE Transactions on Software Engineering, vol. 32, no. 3, pp , Z. Wen and V. Tzerpos, An effectiveness measure for software clustering algorithms, in Proceedings of the 12th IEEE International Workshop on Program Comprehension, 2004, pp

119 Overlap